Научная визуализация, 2026, том 18, номер 1, страницы 20 - 40, DOI: 10.26583/sv.18.1.03
Automated Diabetic Retinopathy Diagnosis and Classification Using Deep Learning with Capsule Network Layers and Stochastic Ensemble Approach
Авторы: M.A Abini1, S Sridevi Sathya Priya2
Department of Electronics and communication Engineering, karunya Institute of technology and sciences, Coimbatore Tamil Nadu, India
1 ORCID: 0009-0005-2419-7770, abinima87@gmail.com
2 ORCID: 0000-0001-9356-6721, sridevi@karunya.edu
Аннотация
Diabetic retinopathy (DR) remains one of the most common vision-related complications of diabetes and requires timely, accurate diagnosis to prevent severe outcomes. Conventional diagnostic approaches rely on the expertise of ophthalmologists, who manually examine retinal images for lesions—a process that can be time-consuming and prone to fatigue-related errors. To address these limitations, this work proposes a fully automated framework for DR detection and stage classification that leverages recent advances in deep learning. The study focuses on the five recognized stages of DR, ranging from the earliest form, non-proliferative diabetic retinopathy (NPDR), through to the advanced proliferative stage (PDR). The method integrates two powerful pre-trained convolutional neural networks, ResNetV2 and MobileNet, with capsule network layers to enhance feature representation. A stochastic ensemble strategy is applied to further strengthen the robustness of predictions. Experimental evaluation on the Kaggle APTOS 2019 dataset demonstrates a test accuracy of 99.81%, outperforming comparable methods in the literature. Performance was assessed using standard metrics such as precision, recall, F1-score, and the ROC curve. Beyond classification accuracy, the approach also offers improved interpretability through capsule-based visualization techniques and ensemble-driven lesion localization, enabling better identification of retinal abnormalities across different DR stages.
Ключевые слова: Diabetic retinopathy, Deep learning, ResNetV2, MobileNet, Capsule networks, stochastic ensemble.
